Sparse-View CT Reconstruction: Methods, Datasets and Where to Start
Last updated:
Sparse-view CT reconstruction recovers an image from projections taken at only a small number of angles. To get started, pair one classical baseline (TV-regularised iterative reconstruction) with one learned method that has released code, and test both on a dataset that ships raw projections: LoDoPaB-CT or the AAPM DL-Sparse-View CT data for simulated sinograms, 2DeteCT or the FleX-ray walnuts for measured ones.
The directory lists 157 sparse-view papers, 40 of them with code. The tables below are generated from those records.
Method families compared
Sparse-view papers in the directory are also tagged with the method family they use, and each family has its own topic page. The table counts the sparse-view papers in each family and names the most recent one with released code. A paper can belong to more than one family.
| Method family | What it is | Papers | With code | Most recent with code |
|---|---|---|---|---|
| Classical & Iterative Reconstruction | Analytical (FBP/FDK) and iterative (SART/SIRT/CGLS) reconstruction, hand-crafted regularizers (TV, TGV, non-local). | 23 | 1 | Plug-and-Play Diffusion Meets ADMM: Dual-Variable Coupling for Robust Medical Image Reconstruction (ICML, 2026) |
| Deep Unrolling & Model-Based Learning | Learned iterative schemes that unroll a classical solver, with or without data consistency. | 22 | 7 | Resolution-Agnostic Neural Operators for Multi-Rate Sparse-View CT (ECCV, 2026) |
| Self-Supervised & Untrained Methods | Training without paired data — deep image prior, noise2noise-style, unsupervised regularization. | 8 | 4 | 3D Field of Junctions: A Noise-Robust, Training-Free Structural Prior for Volumetric Inverse Problems (ECCV, 2026) |
| Diffusion & Generative Priors | Score-based, diffusion and other learned generative priors used as priors or samplers for tomographic inversion. | 27 | 8 | Plug-and-Play Diffusion Meets ADMM: Dual-Variable Coupling for Robust Medical Image Reconstruction (ICML, 2026) |
| Implicit Neural Representations | Coordinate-based networks (INR / NeRF) that represent the volume itself as the unknown. | 11 | 6 | NAB: Neural Adaptive Binning for Sparse-View CT Reconstruction (ICLR, 2026) |
| Gaussian Splatting | Explicit Gaussian primitives as a differentiable tomographic representation. | 9 | 5 | AD-GS: Anatomical Density-guided 3D Gaussian Splatting for Sparse-View Dental CBCT Reconstruction (preprint, 2026) |
Datasets and benchmarks
Benchmarks usually simulate sparse sinograms by keeping a subset of the angles of a full scan, so datasets that ship raw projections are the most useful. Measured datasets test whether a method holds up on real data. Every entry below has its own page with its licence and the papers that use it.
| Name | Data | Modality | Licence |
|---|---|---|---|
| LoDoPaB-CT | Projection + image | 2D fan-beam | CC BY 4.0 |
| 2DeteCT | Projection + image | 2D fan-beam | CC BY 4.0 |
| AAPM DL-Sparse-View CT data | Projection + image | 2D fan-beam (simulated) | CC BY 4.0 |
| SparseBeads | Projection + image | micro-CT | CC BY 4.0 |
| FleX-ray Walnut CBCT Collection | Projection + image | cone-beam micro-CT | CC BY 4.0 |
| Carved cheese (FIPS) | Projection only | 2D fan-beam | CC BY 4.0 |
| Name | Year | What it evaluates |
|---|---|---|
| AAPM DL-Sparse-View CT Challenge | 2021 | Recover a 2D breast-CT phantom from ideal noiseless 128-view fan-beam data: 4,000 training pairs, 100 test cases, ranked by RMSE on a public leaderboard. |
Where to start
These are the maintainers' start-here picks from the sparse-view CT topic page, from the classical TV baseline to per-scan neural fields and Gaussian splatting. For data with a contiguous missing wedge rather than spread-out gaps, see limited-angle tomography instead.
| Paper | Venue | Year | Why read it |
|---|---|---|---|
| Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization | PMB | 2008 | The TV-constrained baseline that most sparse-view papers still compare against. |
| A Sparse-View CT Reconstruction Method Based on Combination of DenseNet and Deconvolution · code | TMI | 2018 | An early, simple CNN that post-processes the FBP image, with code; a useful first deep-learning baseline. |
| DRONE: Dual-Domain Residual-Based Optimization Network for Sparse-View CT Reconstruction · code | TMI | 2021 | A dual-domain network that refines both the sinogram and the image while enforcing consistency, with code. |
| NAF: Neural Attenuation Fields for Sparse-View CBCT Reconstruction · code | MICCAI | 2022 | A self-supervised neural attenuation field for sparse-view CBCT that fits each scan without training data, with code. |
| R2-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction · code | NeurIPS | 2024 | The first Gaussian-splatting framework for sparse-view tomographic reconstruction; fast, with code. |
Related pages
- Sparse-View CT
- Classical & Iterative Reconstruction
- Deep Unrolling & Model-Based Learning
- Self-Supervised & Untrained Methods
- Diffusion & Generative Priors
- Implicit Neural Representations
- Gaussian Splatting
- CT Reconstruction Datasets and Benchmarks
- Low-Dose CT Denoising: Benchmarks, Datasets and Toolkits
- How to Choose a CT Reconstruction Dataset
Frequently asked questions
What is sparse-view CT?
Sparse-view CT reconstructs an image from projections taken at only a small number of angles, typically tens rather than the hundreds or thousands of views in a standard scan. Fewer views mean lower dose and faster acquisition, but filtered back-projection then produces strong streak artifacts.
How is sparse-view CT different from limited-angle CT?
In sparse-view CT the missing views are gaps spread across the scan. In limited-angle CT they are contiguous, a missing wedge of directions, so edges whose orientation falls inside the wedge are essentially invisible and the problem is more severely ill-posed. Each has its own topic page.
Which dataset should I use for sparse-view CT?
For simulated sparse sinograms, LoDoPaB-CT and the AAPM DL-Sparse-View CT data; the latter isolates the sparse-view problem with noiseless 128-view sinograms. For measured data, use 2DeteCT, the FleX-ray walnut collection or SparseBeads.
Do I need training data for sparse-view reconstruction?
Not always. Per-scan representations such as implicit neural fields and Gaussian splatting are fitted to each scan, and self-supervised methods train without paired data. NAF, for example, is a self-supervised neural attenuation field for sparse-view CBCT that fits each scan without training data.